Skip to content
InfoResearchPreprint

As Student Rollouts Continue, Visual Backdoors Keep Growing

Published
Record updated
View JSON

Summary

This paper studies how visual backdoors propagate from teacher to student models during On-Policy Distillation (OPD) for Vision-Language Models (VLMs). Experiments across multiple model families and datasets, plus theoretical analysis, show backdoors rarely emerge at low poisoning ratios but reach nearly 100% attack success rate once a critical threshold is exceeded. The authors attribute propagation to the continuous accumulation of anomalous probability mass from the teacher, and propose a trigger design that enables propagation with zero explicit trigger injection.